MetaCon: Revitalizing Internet Congestion Control with Meta-Reinforcement Learning
He Bai, Hui Li, Jianming Que, Minglong Zhang, Peter Han Joo Chong, Kalupahana Liyanage Kushan Sudheera, Xinyuan Pei
Abstract
Effective congestion control algorithms (CCAs) are crucial for the smooth operation of Internet communication infrastructure. CCAs adjust transmission rates based on congestion signals, optimizing resource utilization and user experience. However, existing studies, both rule-based and learning-based CCAs, often struggle with generalization and underperform when deployed in real-world environments. When applied to unseen network conditions, hand-crafted schemes or pre-trained models may experience significant performance degradation. To address this challenge, we propose MetaCon, a novel adaptive Internet congestion control approach based on meta-reinforcement learning. MetaCon leverages knowledge learned from prior scenarios to quickly adapt to new environments. Experimental results show that MetaCon outperforms existing algorithms by exhibiting superior generalization and achieving better transmission performance across a wide variety of network conditions.
BibTeX
@inproceedings{icassp2025_metaconrevitaliz,
title = {MetaCon: Revitalizing Internet Congestion Control with Meta-Reinforcement Learning},
author = {He Bai and Hui Li and Jianming Que and Minglong Zhang and Peter Han Joo Chong and Kalupahana Liyanage Kushan Sudheera and Xinyuan Pei},
booktitle = {ICASSP 2025},
year = {2025}
}